2 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
3 * The Regents of the University of California. All rights reserved.
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
13 * 4. Neither the name of the University nor the names of its contributors
14 * may be used to endorse or promote products derived from this software
15 * without specific prior written permission.
17 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29 * @(#)tcp_output.c 8.4 (Berkeley) 5/24/95
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
36 #include "opt_inet6.h"
37 #include "opt_ipsec.h"
39 #include "opt_tcpdebug.h"
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/domain.h>
44 #include <sys/kernel.h>
47 #include <sys/mutex.h>
48 #include <sys/protosw.h>
49 #include <sys/socket.h>
50 #include <sys/socketvar.h>
51 #include <sys/sysctl.h>
52 #include <sys/vimage.h>
54 #include <net/route.h>
56 #include <netinet/in.h>
57 #include <netinet/in_systm.h>
58 #include <netinet/ip.h>
59 #include <netinet/in_pcb.h>
60 #include <netinet/ip_var.h>
61 #include <netinet/ip_options.h>
63 #include <netinet6/in6_pcb.h>
64 #include <netinet/ip6.h>
65 #include <netinet6/ip6_var.h>
67 #include <netinet/tcp.h>
69 #include <netinet/tcp_fsm.h>
70 #include <netinet/tcp_seq.h>
71 #include <netinet/tcp_timer.h>
72 #include <netinet/tcp_var.h>
73 #include <netinet/tcpip.h>
75 #include <netinet/tcp_debug.h>
79 #include <netipsec/ipsec.h>
82 #include <machine/in_cksum.h>
84 #include <security/mac/mac_framework.h>
87 extern struct mbuf *m_copypack();
90 int path_mtu_discovery = 1;
91 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
92 &path_mtu_discovery, 1, "Enable Path MTU Discovery");
95 SYSCTL_INT(_net_inet_tcp, OID_AUTO, slowstart_flightsize, CTLFLAG_RW,
96 &ss_fltsz, 1, "Slow start flight size");
98 int ss_fltsz_local = 4;
99 SYSCTL_INT(_net_inet_tcp, OID_AUTO, local_slowstart_flightsize, CTLFLAG_RW,
100 &ss_fltsz_local, 1, "Slow start flight size for local networks");
102 int tcp_do_newreno = 1;
103 SYSCTL_INT(_net_inet_tcp, OID_AUTO, newreno, CTLFLAG_RW,
104 &tcp_do_newreno, 0, "Enable NewReno Algorithms");
107 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW,
108 &tcp_do_tso, 0, "Enable TCP Segmentation Offload");
110 int tcp_do_autosndbuf = 1;
111 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW,
112 &tcp_do_autosndbuf, 0, "Enable automatic send buffer sizing");
114 int tcp_autosndbuf_inc = 8*1024;
115 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW,
116 &tcp_autosndbuf_inc, 0, "Incrementor step size of automatic send buffer");
118 int tcp_autosndbuf_max = 256*1024;
119 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW,
120 &tcp_autosndbuf_max, 0, "Max size of automatic send buffer");
124 * Tcp output routine: figure out what should be sent and send it.
127 tcp_output(struct tcpcb *tp)
129 struct socket *so = tp->t_inpcb->inp_socket;
130 long len, recwin, sendwin;
131 int off, flags, error;
133 struct ip *ip = NULL;
134 struct ipovly *ipov = NULL;
136 u_char opt[TCP_MAXOLEN];
137 unsigned ipoptlen, optlen, hdrlen;
139 unsigned ipsec_optlen = 0;
142 int sack_rxmit, sack_bytes_rxmt;
147 int maxburst = TCP_MAXBURST;
150 struct ip6_hdr *ip6 = NULL;
153 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
156 INP_WLOCK_ASSERT(tp->t_inpcb);
159 * Determine length of data that should be transmitted,
160 * and flags that will be used.
161 * If there is some data or critical controls (SYN, RST)
162 * to send, then transmit; otherwise, investigate further.
164 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
165 if (idle && (ticks - tp->t_rcvtime) >= tp->t_rxtcur) {
167 * We have been idle for "a while" and no acks are
168 * expected to clock out any data we send --
169 * slow start to get ack "clock" running again.
171 * Set the slow-start flight size depending on whether
172 * this is a local network or not.
177 if (in6_localaddr(&tp->t_inpcb->in6p_faddr))
178 ss = V_ss_fltsz_local;
181 if (in_localaddr(tp->t_inpcb->inp_faddr))
182 ss = V_ss_fltsz_local;
183 tp->snd_cwnd = tp->t_maxseg * ss;
185 tp->t_flags &= ~TF_LASTIDLE;
187 if (tp->t_flags & TF_MORETOCOME) {
188 tp->t_flags |= TF_LASTIDLE;
194 * If we've recently taken a timeout, snd_max will be greater than
195 * snd_nxt. There may be SACK information that allows us to avoid
196 * resending already delivered data. Adjust snd_nxt accordingly.
198 if ((tp->t_flags & TF_SACK_PERMIT) &&
199 SEQ_LT(tp->snd_nxt, tp->snd_max))
202 off = tp->snd_nxt - tp->snd_una;
203 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
204 sendwin = min(sendwin, tp->snd_bwnd);
206 flags = tcp_outflags[tp->t_state];
208 * Send any SACK-generated retransmissions. If we're explicitly trying
209 * to send out new data (when sendalot is 1), bypass this function.
210 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
211 * we're replacing a (future) new transmission with a retransmission
212 * now, and we previously incremented snd_cwnd in tcp_input().
215 * Still in sack recovery , reset rxmit flag to zero.
221 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp) &&
222 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
225 cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
228 /* Do not retransmit SACK segments beyond snd_recover */
229 if (SEQ_GT(p->end, tp->snd_recover)) {
231 * (At least) part of sack hole extends beyond
232 * snd_recover. Check to see if we can rexmit data
235 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
237 * Can't rexmit any more data for this hole.
238 * That data will be rexmitted in the next
239 * sack recovery episode, when snd_recover
240 * moves past p->rxmit.
243 goto after_sack_rexmit;
245 /* Can rexmit part of the current hole */
246 len = ((long)ulmin(cwin,
247 tp->snd_recover - p->rxmit));
249 len = ((long)ulmin(cwin, p->end - p->rxmit));
250 off = p->rxmit - tp->snd_una;
251 KASSERT(off >= 0,("%s: sack block to the left of una : %d",
256 V_tcpstat.tcps_sack_rexmits++;
257 V_tcpstat.tcps_sack_rexmit_bytes +=
258 min(len, tp->t_maxseg);
263 * Get standard flags, and add SYN or FIN if requested by 'hidden'
266 if (tp->t_flags & TF_NEEDFIN)
268 if (tp->t_flags & TF_NEEDSYN)
271 SOCKBUF_LOCK(&so->so_snd);
273 * If in persist timeout with window of 0, send 1 byte.
274 * Otherwise, if window is small but nonzero
275 * and timer expired, we will send what we can
276 * and go to transmit state.
278 if (tp->t_flags & TF_FORCEDATA) {
281 * If we still have some data to send, then
282 * clear the FIN bit. Usually this would
283 * happen below when it realizes that we
284 * aren't sending all the data. However,
285 * if we have exactly 1 byte of unsent data,
286 * then it won't clear the FIN bit below,
287 * and if we are in persist state, we wind
288 * up sending the packet without recording
289 * that we sent the FIN bit.
291 * We can't just blindly clear the FIN bit,
292 * because if we don't have any more data
293 * to send then the probe will be the FIN
296 if (off < so->so_snd.sb_cc)
300 tcp_timer_activate(tp, TT_PERSIST, 0);
306 * If snd_nxt == snd_max and we have transmitted a FIN, the
307 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
308 * a negative length. This can also occur when TCP opens up
309 * its congestion window while receiving additional duplicate
310 * acks after fast-retransmit because TCP will reset snd_nxt
311 * to snd_max after the fast-retransmit.
313 * In the normal retransmit-FIN-only case, however, snd_nxt will
314 * be set to snd_una, the offset will be 0, and the length may
317 * If sack_rxmit is true we are retransmitting from the scoreboard
318 * in which case len is already set.
320 if (sack_rxmit == 0) {
321 if (sack_bytes_rxmt == 0)
322 len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off);
327 * We are inside of a SACK recovery episode and are
328 * sending new data, having retransmitted all the
329 * data possible in the scoreboard.
331 len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd)
334 * Don't remove this (len > 0) check !
335 * We explicitly check for len > 0 here (although it
336 * isn't really necessary), to work around a gcc
337 * optimization issue - to force gcc to compute
338 * len above. Without this check, the computation
339 * of len is bungled by the optimizer.
342 cwin = tp->snd_cwnd -
343 (tp->snd_nxt - tp->sack_newdata) -
347 len = lmin(len, cwin);
353 * Lop off SYN bit if it has already been sent. However, if this
354 * is SYN-SENT state and if segment contains data and if we don't
355 * know that foreign host supports TAO, suppress sending segment.
357 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
358 if (tp->t_state != TCPS_SYN_RECEIVED)
364 * Be careful not to send data and/or FIN on SYN segments.
365 * This measure is needed to prevent interoperability problems
366 * with not fully conformant TCP implementations.
368 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
375 * If FIN has been sent but not acked,
376 * but we haven't been called to retransmit,
377 * len will be < 0. Otherwise, window shrank
378 * after we sent into it. If window shrank to 0,
379 * cancel pending retransmit, pull snd_nxt back
380 * to (closed) window, and set the persist timer
381 * if it isn't already going. If the window didn't
382 * close completely, just wait for an ACK.
386 tcp_timer_activate(tp, TT_REXMT, 0);
388 tp->snd_nxt = tp->snd_una;
389 if (!tcp_timer_active(tp, TT_PERSIST))
394 /* len will be >= 0 after this point. */
395 KASSERT(len >= 0, ("%s: len < 0", __func__));
398 * Automatic sizing of send socket buffer. Often the send buffer
399 * size is not optimally adjusted to the actual network conditions
400 * at hand (delay bandwidth product). Setting the buffer size too
401 * small limits throughput on links with high bandwidth and high
402 * delay (eg. trans-continental/oceanic links). Setting the
403 * buffer size too big consumes too much real kernel memory,
404 * especially with many connections on busy servers.
406 * The criteria to step up the send buffer one notch are:
407 * 1. receive window of remote host is larger than send buffer
408 * (with a fudge factor of 5/4th);
409 * 2. send buffer is filled to 7/8th with data (so we actually
410 * have data to make use of it);
411 * 3. send buffer fill has not hit maximal automatic size;
412 * 4. our send window (slow start and cogestion controlled) is
413 * larger than sent but unacknowledged data in send buffer.
415 * The remote host receive window scaling factor may limit the
416 * growing of the send buffer before it reaches its allowed
419 * It scales directly with slow start or congestion window
420 * and does at most one step per received ACK. This fast
421 * scaling has the drawback of growing the send buffer beyond
422 * what is strictly necessary to make full use of a given
423 * delay*bandwith product. However testing has shown this not
424 * to be much of an problem. At worst we are trading wasting
425 * of available bandwith (the non-use of it) for wasting some
426 * socket buffer memory.
428 * TODO: Shrink send buffer during idle periods together
429 * with congestion window. Requires another timer. Has to
430 * wait for upcoming tcp timer rewrite.
432 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
433 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
434 so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
435 so->so_snd.sb_cc < V_tcp_autosndbuf_max &&
436 sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
437 if (!sbreserve_locked(&so->so_snd,
438 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
439 V_tcp_autosndbuf_max), so, curthread))
440 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
445 * Truncate to the maximum segment length or enable TCP Segmentation
446 * Offloading (if supported by hardware) and ensure that FIN is removed
447 * if the length no longer contains the last data byte.
449 * TSO may only be used if we are in a pure bulk sending state. The
450 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
451 * IP options prevent using TSO. With TSO the TCP header is the same
452 * (except for the sequence number) for all generated packets. This
453 * makes it impossible to transmit any options which vary per generated
456 * The length of TSO bursts is limited to TCP_MAXWIN. That limit and
457 * removal of FIN (if not already catched here) are handled later after
458 * the exact length of the TCP options are known.
462 * Pre-calculate here as we save another lookup into the darknesses
463 * of IPsec that way and can actually decide if TSO is ok.
465 ipsec_optlen = ipsec_hdrsiz_tcp(tp);
467 if (len > tp->t_maxseg) {
468 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso &&
469 ((tp->t_flags & TF_SIGNATURE) == 0) &&
470 tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
471 tp->t_inpcb->inp_options == NULL &&
472 tp->t_inpcb->in6p_options == NULL
485 if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
488 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
492 recwin = sbspace(&so->so_rcv);
495 * Sender silly window avoidance. We transmit under the following
496 * conditions when len is non-zero:
498 * - We have a full segment (or more with TSO)
499 * - This is the last buffer in a write()/send() and we are
500 * either idle or running NODELAY
501 * - we've timed out (e.g. persist timer)
502 * - we have more then 1/2 the maximum send window's worth of
503 * data (receiver may be limited the window size)
504 * - we need to retransmit
507 if (len >= tp->t_maxseg)
510 * NOTE! on localhost connections an 'ack' from the remote
511 * end may occur synchronously with the output and cause
512 * us to flush a buffer queued with moretocome. XXX
514 * note: the len + off check is almost certainly unnecessary.
516 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
517 (idle || (tp->t_flags & TF_NODELAY)) &&
518 len + off >= so->so_snd.sb_cc &&
519 (tp->t_flags & TF_NOPUSH) == 0) {
522 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */
524 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
526 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
533 * Compare available window to amount of window
534 * known to peer (as advertised window less
535 * next expected input). If the difference is at least two
536 * max size segments, or at least 50% of the maximum possible
537 * window, then want to send a window update to peer.
538 * Skip this if the connection is in T/TCP half-open state.
539 * Don't send pure window updates when the peer has closed
540 * the connection and won't ever send more data.
542 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
543 !TCPS_HAVERCVDFIN(tp->t_state)) {
545 * "adv" is the amount we can increase the window,
546 * taking into account that we are limited by
547 * TCP_MAXWIN << tp->rcv_scale.
549 long adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale) -
550 (tp->rcv_adv - tp->rcv_nxt);
552 if (adv >= (long) (2 * tp->t_maxseg))
554 if (2 * adv >= (long) so->so_rcv.sb_hiwat)
559 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
560 * is also a catch-all for the retransmit timer timeout case.
562 if (tp->t_flags & TF_ACKNOW)
564 if ((flags & TH_RST) ||
565 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
567 if (SEQ_GT(tp->snd_up, tp->snd_una))
570 * If our state indicates that FIN should be sent
571 * and we have not yet done so, then we need to send.
573 if (flags & TH_FIN &&
574 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
577 * In SACK, it is possible for tcp_output to fail to send a segment
578 * after the retransmission timer has been turned off. Make sure
579 * that the retransmission timer is set.
581 if ((tp->t_flags & TF_SACK_PERMIT) &&
582 SEQ_GT(tp->snd_max, tp->snd_una) &&
583 !tcp_timer_active(tp, TT_REXMT) &&
584 !tcp_timer_active(tp, TT_PERSIST)) {
585 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
589 * TCP window updates are not reliable, rather a polling protocol
590 * using ``persist'' packets is used to insure receipt of window
591 * updates. The three ``states'' for the output side are:
592 * idle not doing retransmits or persists
593 * persisting to move a small or zero window
594 * (re)transmitting and thereby not persisting
596 * tcp_timer_active(tp, TT_PERSIST)
597 * is true when we are in persist state.
598 * (tp->t_flags & TF_FORCEDATA)
599 * is set when we are called to send a persist packet.
600 * tcp_timer_active(tp, TT_REXMT)
601 * is set when we are retransmitting
602 * The output side is idle when both timers are zero.
604 * If send window is too small, there is data to transmit, and no
605 * retransmit or persist is pending, then go to persist state.
606 * If nothing happens soon, send when timer expires:
607 * if window is nonzero, transmit what we can,
608 * otherwise force out a byte.
610 if (so->so_snd.sb_cc && !tcp_timer_active(tp, TT_REXMT) &&
611 !tcp_timer_active(tp, TT_PERSIST)) {
617 * No reason to send a segment, just return.
620 SOCKBUF_UNLOCK(&so->so_snd);
624 SOCKBUF_LOCK_ASSERT(&so->so_snd);
626 * Before ESTABLISHED, force sending of initial options
627 * unless TCP set not to do any options.
628 * NOTE: we assume that the IP/TCP header plus TCP options
629 * always fit in a single mbuf, leaving room for a maximum
631 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
636 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
639 hdrlen = sizeof (struct tcpiphdr);
642 * Compute options for segment.
643 * We only have to care about SYN and established connection
644 * segments. Options for SYN-ACK segments are handled in TCP
647 if ((tp->t_flags & TF_NOOPT) == 0) {
649 /* Maximum segment size. */
650 if (flags & TH_SYN) {
651 tp->snd_nxt = tp->iss;
652 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
653 to.to_flags |= TOF_MSS;
655 /* Window scaling. */
656 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
657 to.to_wscale = tp->request_r_scale;
658 to.to_flags |= TOF_SCALE;
661 if ((tp->t_flags & TF_RCVD_TSTMP) ||
662 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
663 to.to_tsval = ticks + tp->ts_offset;
664 to.to_tsecr = tp->ts_recent;
665 to.to_flags |= TOF_TS;
666 /* Set receive buffer autosizing timestamp. */
667 if (tp->rfbuf_ts == 0 &&
668 (so->so_rcv.sb_flags & SB_AUTOSIZE))
669 tp->rfbuf_ts = ticks;
671 /* Selective ACK's. */
672 if (tp->t_flags & TF_SACK_PERMIT) {
674 to.to_flags |= TOF_SACKPERM;
675 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
676 (tp->t_flags & TF_SACK_PERMIT) &&
677 tp->rcv_numsacks > 0) {
678 to.to_flags |= TOF_SACK;
679 to.to_nsacks = tp->rcv_numsacks;
680 to.to_sacks = (u_char *)tp->sackblks;
684 /* TCP-MD5 (RFC2385). */
686 if (!isipv6 && (tp->t_flags & TF_SIGNATURE))
688 if (tp->t_flags & TF_SIGNATURE)
690 to.to_flags |= TOF_SIGNATURE;
691 #endif /* TCP_SIGNATURE */
693 /* Processing the options. */
694 hdrlen += optlen = tcp_addoptions(&to, opt);
699 ipoptlen = ip6_optlen(tp->t_inpcb);
702 if (tp->t_inpcb->inp_options)
703 ipoptlen = tp->t_inpcb->inp_options->m_len -
704 offsetof(struct ipoption, ipopt_list);
708 ipoptlen += ipsec_optlen;
712 * Adjust data length if insertion of options will
713 * bump the packet length beyond the t_maxopd length.
714 * Clear the FIN bit because we cut off the tail of
717 * When doing TSO limit a burst to TCP_MAXWIN minus the
718 * IP, TCP and Options length to keep ip->ip_len from
719 * overflowing. Prevent the last segment from being
720 * fractional thus making them all equal sized and set
721 * the flag to continue sending. TSO is disabled when
722 * IP options or IPSEC are present.
724 if (len + optlen + ipoptlen > tp->t_maxopd) {
727 if (len > TCP_MAXWIN - hdrlen - optlen) {
728 len = TCP_MAXWIN - hdrlen - optlen;
729 len = len - (len % (tp->t_maxopd - optlen));
731 } else if (tp->t_flags & TF_NEEDFIN)
734 len = tp->t_maxopd - optlen - ipoptlen;
739 /*#ifdef DIAGNOSTIC*/
741 if (max_linkhdr + hdrlen > MCLBYTES)
743 if (max_linkhdr + hdrlen > MHLEN)
745 panic("tcphdr too big");
749 * Grab a header mbuf, attaching a copy of data to
750 * be transmitted, and initialize the header from
751 * the template for sends on this connection.
757 if ((tp->t_flags & TF_FORCEDATA) && len == 1)
758 V_tcpstat.tcps_sndprobe++;
759 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
760 V_tcpstat.tcps_sndrexmitpack++;
761 V_tcpstat.tcps_sndrexmitbyte += len;
763 V_tcpstat.tcps_sndpack++;
764 V_tcpstat.tcps_sndbyte += len;
767 if ((m = m_copypack(so->so_snd.sb_mb, off,
768 (int)len, max_linkhdr + hdrlen)) == 0) {
769 SOCKBUF_UNLOCK(&so->so_snd);
774 * m_copypack left space for our hdr; use it.
779 MGETHDR(m, M_DONTWAIT, MT_DATA);
781 SOCKBUF_UNLOCK(&so->so_snd);
786 if (MHLEN < hdrlen + max_linkhdr) {
787 MCLGET(m, M_DONTWAIT);
788 if ((m->m_flags & M_EXT) == 0) {
789 SOCKBUF_UNLOCK(&so->so_snd);
796 m->m_data += max_linkhdr;
800 * Start the m_copy functions from the closest mbuf
801 * to the offset in the socket buffer chain.
803 mb = sbsndptr(&so->so_snd, off, len, &moff);
805 if (len <= MHLEN - hdrlen - max_linkhdr) {
806 m_copydata(mb, moff, (int)len,
807 mtod(m, caddr_t) + hdrlen);
810 m->m_next = m_copy(mb, moff, (int)len);
811 if (m->m_next == NULL) {
812 SOCKBUF_UNLOCK(&so->so_snd);
820 * If we're sending everything we've got, set PUSH.
821 * (This will keep happy those implementations which only
822 * give data to the user when a buffer fills or
825 if (off + len == so->so_snd.sb_cc)
827 SOCKBUF_UNLOCK(&so->so_snd);
829 SOCKBUF_UNLOCK(&so->so_snd);
830 if (tp->t_flags & TF_ACKNOW)
831 V_tcpstat.tcps_sndacks++;
832 else if (flags & (TH_SYN|TH_FIN|TH_RST))
833 V_tcpstat.tcps_sndctrl++;
834 else if (SEQ_GT(tp->snd_up, tp->snd_una))
835 V_tcpstat.tcps_sndurg++;
837 V_tcpstat.tcps_sndwinup++;
839 MGETHDR(m, M_DONTWAIT, MT_DATA);
845 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
850 m->m_data += max_linkhdr;
853 SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
854 m->m_pkthdr.rcvif = (struct ifnet *)0;
856 mac_inpcb_create_mbuf(tp->t_inpcb, m);
860 ip6 = mtod(m, struct ip6_hdr *);
861 th = (struct tcphdr *)(ip6 + 1);
862 tcpip_fillheaders(tp->t_inpcb, ip6, th);
866 ip = mtod(m, struct ip *);
867 ipov = (struct ipovly *)ip;
868 th = (struct tcphdr *)(ip + 1);
869 tcpip_fillheaders(tp->t_inpcb, ip, th);
873 * Fill in fields, remembering maximum advertised
874 * window for use in delaying messages about window sizes.
875 * If resending a FIN, be sure not to use a new sequence number.
877 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
878 tp->snd_nxt == tp->snd_max)
881 * If we are starting a connection, send ECN setup
882 * SYN packet. If we are on a retransmit, we may
883 * resend those bits a number of times as per
886 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
887 if (tp->t_rxtshift >= 1) {
888 if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
889 flags |= TH_ECE|TH_CWR;
891 flags |= TH_ECE|TH_CWR;
894 if (tp->t_state == TCPS_ESTABLISHED &&
895 (tp->t_flags & TF_ECN_PERMIT)) {
897 * If the peer has ECN, mark data packets with
898 * ECN capable transmission (ECT).
899 * Ignore pure ack packets, retransmissions and window probes.
901 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
902 !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
905 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
908 ip->ip_tos |= IPTOS_ECN_ECT0;
909 V_tcpstat.tcps_ecn_ect0++;
913 * Reply with proper ECN notifications.
915 if (tp->t_flags & TF_ECN_SND_CWR) {
917 tp->t_flags &= ~TF_ECN_SND_CWR;
919 if (tp->t_flags & TF_ECN_SND_ECE)
924 * If we are doing retransmissions, then snd_nxt will
925 * not reflect the first unsent octet. For ACK only
926 * packets, we do not want the sequence number of the
927 * retransmitted packet, we want the sequence number
928 * of the next unsent octet. So, if there is no data
929 * (and no SYN or FIN), use snd_max instead of snd_nxt
930 * when filling in ti_seq. But if we are in persist
931 * state, snd_max might reflect one byte beyond the
932 * right edge of the window, so use snd_nxt in that
933 * case, since we know we aren't doing a retransmission.
934 * (retransmit and persist are mutually exclusive...)
936 if (sack_rxmit == 0) {
937 if (len || (flags & (TH_SYN|TH_FIN)) ||
938 tcp_timer_active(tp, TT_PERSIST))
939 th->th_seq = htonl(tp->snd_nxt);
941 th->th_seq = htonl(tp->snd_max);
943 th->th_seq = htonl(p->rxmit);
945 tp->sackhint.sack_bytes_rexmit += len;
947 th->th_ack = htonl(tp->rcv_nxt);
949 bcopy(opt, th + 1, optlen);
950 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
952 th->th_flags = flags;
954 * Calculate receive window. Don't shrink window,
955 * but avoid silly window syndrome.
957 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
958 recwin < (long)tp->t_maxseg)
960 if (recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
961 recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
962 if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
963 recwin = (long)TCP_MAXWIN << tp->rcv_scale;
966 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
967 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK>
968 * case is handled in syncache.
971 th->th_win = htons((u_short)
972 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
974 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
977 * Adjust the RXWIN0SENT flag - indicate that we have advertised
978 * a 0 window. This may cause the remote transmitter to stall. This
979 * flag tells soreceive() to disable delayed acknowledgements when
980 * draining the buffer. This can occur if the receiver is attempting
981 * to read more data than can be buffered prior to transmitting on
985 tp->t_flags |= TF_RXWIN0SENT;
987 tp->t_flags &= ~TF_RXWIN0SENT;
988 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
989 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
990 th->th_flags |= TH_URG;
993 * If no urgent pointer to send, then we pull
994 * the urgent pointer to the left edge of the send window
995 * so that it doesn't drift into the send window on sequence
998 tp->snd_up = tp->snd_una; /* drag it along */
1000 #ifdef TCP_SIGNATURE
1004 if (tp->t_flags & TF_SIGNATURE) {
1005 int sigoff = to.to_signature - opt;
1006 tcp_signature_compute(m, sizeof(struct ip), len, optlen,
1007 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
1012 * Put TCP length in extended header, and then
1013 * checksum extended header and data.
1015 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1019 * ip6_plen is not need to be filled now, and will be filled
1022 th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
1023 sizeof(struct tcphdr) + optlen + len);
1027 m->m_pkthdr.csum_flags = CSUM_TCP;
1028 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1029 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1030 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1032 /* IP version must be set here for ipv4/ipv6 checking later */
1033 KASSERT(ip->ip_v == IPVERSION,
1034 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1038 * Enable TSO and specify the size of the segments.
1039 * The TCP pseudo header checksum is always provided.
1040 * XXX: Fixme: This is currently not the case for IPv6.
1043 m->m_pkthdr.csum_flags = CSUM_TSO;
1044 m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen;
1048 * In transmit state, time the transmission and arrange for
1049 * the retransmit. In persist state, just set snd_max.
1051 if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1052 !tcp_timer_active(tp, TT_PERSIST)) {
1053 tcp_seq startseq = tp->snd_nxt;
1056 * Advance snd_nxt over sequence space of this segment.
1058 if (flags & (TH_SYN|TH_FIN)) {
1061 if (flags & TH_FIN) {
1063 tp->t_flags |= TF_SENTFIN;
1069 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1070 tp->snd_max = tp->snd_nxt;
1072 * Time this transmission if not a retransmission and
1073 * not currently timing anything.
1075 if (tp->t_rtttime == 0) {
1076 tp->t_rtttime = ticks;
1077 tp->t_rtseq = startseq;
1078 V_tcpstat.tcps_segstimed++;
1083 * Set retransmit timer if not currently set,
1084 * and not doing a pure ack or a keep-alive probe.
1085 * Initial value for retransmit timer is smoothed
1086 * round-trip time + 2 * round-trip time variance.
1087 * Initialize shift counter which is used for backoff
1088 * of retransmit time.
1091 if (!tcp_timer_active(tp, TT_REXMT) &&
1092 ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1093 (tp->snd_nxt != tp->snd_una))) {
1094 if (tcp_timer_active(tp, TT_PERSIST)) {
1095 tcp_timer_activate(tp, TT_PERSIST, 0);
1098 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1102 * Persist case, update snd_max but since we are in
1103 * persist mode (no window) we do not update snd_nxt.
1108 if (flags & TH_FIN) {
1110 tp->t_flags |= TF_SENTFIN;
1112 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1113 tp->snd_max = tp->snd_nxt + len;
1120 if (so->so_options & SO_DEBUG) {
1126 save = ipov->ih_len;
1127 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1129 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1133 ipov->ih_len = save;
1138 * Fill in IP length and desired time to live and
1139 * send to IP level. There should be a better way
1140 * to handle ttl and tos; we could keep them in
1141 * the template, but need a way to checksum without them.
1144 * m->m_pkthdr.len should have been set before cksum calcuration,
1145 * because in6_cksum() need it.
1150 * we separately set hoplimit for every segment, since the
1151 * user might want to change the value via setsockopt.
1152 * Also, desired default hop limit might be changed via
1153 * Neighbor Discovery.
1155 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1157 /* TODO: IPv6 IP6TOS_ECT bit on */
1158 error = ip6_output(m,
1159 tp->t_inpcb->in6p_outputopts, NULL,
1160 ((so->so_options & SO_DONTROUTE) ?
1161 IP_ROUTETOIF : 0), NULL, NULL, tp->t_inpcb);
1165 ip->ip_len = m->m_pkthdr.len;
1167 if (INP_CHECK_SOCKAF(so, AF_INET6))
1168 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1171 * If we do path MTU discovery, then we set DF on every packet.
1172 * This might not be the best thing to do according to RFC3390
1173 * Section 2. However the tcp hostcache migitates the problem
1174 * so it affects only the first tcp connection with a host.
1176 if (V_path_mtu_discovery)
1177 ip->ip_off |= IP_DF;
1179 error = ip_output(m, tp->t_inpcb->inp_options, NULL,
1180 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1186 * We know that the packet was lost, so back out the
1187 * sequence number advance, if any.
1189 * If the error is EPERM the packet got blocked by the
1190 * local firewall. Normally we should terminate the
1191 * connection but the blocking may have been spurious
1192 * due to a firewall reconfiguration cycle. So we treat
1193 * it like a packet loss and let the retransmit timer and
1194 * timeouts do their work over time.
1195 * XXX: It is a POLA question whether calling tcp_drop right
1196 * away would be the really correct behavior instead.
1198 if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1199 !tcp_timer_active(tp, TT_PERSIST)) &&
1200 ((flags & TH_SYN) == 0) &&
1204 tp->sackhint.sack_bytes_rexmit -= len;
1205 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1206 ("sackhint bytes rtx >= 0"));
1211 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */
1214 tp->t_softerror = error;
1217 if (!tcp_timer_active(tp, TT_REXMT) &&
1218 !tcp_timer_active(tp, TT_PERSIST))
1219 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1220 tp->snd_cwnd = tp->t_maxseg;
1224 * For some reason the interface we used initially
1225 * to send segments changed to another or lowered
1228 * tcp_mtudisc() will find out the new MTU and as
1229 * its last action, initiate retransmission, so it
1230 * is important to not do so here.
1232 * If TSO was active we either got an interface
1233 * without TSO capabilits or TSO was turned off.
1234 * Disable it for this connection as too and
1235 * immediatly retry with MSS sized segments generated
1239 tp->t_flags &= ~TF_TSO;
1240 tcp_mtudisc(tp->t_inpcb, 0);
1246 if (TCPS_HAVERCVDSYN(tp->t_state)) {
1247 tp->t_softerror = error;
1255 V_tcpstat.tcps_sndtotal++;
1258 * Data sent (as far as we can tell).
1259 * If this advertises a larger window than any other segment,
1260 * then remember the size of the advertised window.
1261 * Any pending ACK has now been sent.
1263 if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1264 tp->rcv_adv = tp->rcv_nxt + recwin;
1265 tp->last_ack_sent = tp->rcv_nxt;
1266 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1267 if (tcp_timer_active(tp, TT_DELACK))
1268 tcp_timer_activate(tp, TT_DELACK, 0);
1271 * This completely breaks TCP if newreno is turned on. What happens
1272 * is that if delayed-acks are turned on on the receiver, this code
1273 * on the transmitter effectively destroys the TCP window, forcing
1274 * it to four packets (1.5Kx4 = 6K window).
1276 if (sendalot && (!V_tcp_do_newreno || --maxburst))
1285 tcp_setpersist(struct tcpcb *tp)
1287 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1290 if (tcp_timer_active(tp, TT_REXMT))
1291 panic("tcp_setpersist: retransmit pending");
1293 * Start/restart persistance timer.
1295 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1296 TCPTV_PERSMIN, TCPTV_PERSMAX);
1297 tcp_timer_activate(tp, TT_PERSIST, tt);
1298 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1303 * Insert TCP options according to the supplied parameters to the place
1304 * optp in a consistent way. Can handle unaligned destinations.
1306 * The order of the option processing is crucial for optimal packing and
1307 * alignment for the scarce option space.
1309 * The optimal order for a SYN/SYN-ACK segment is:
1310 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1311 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1313 * The SACK options should be last. SACK blocks consume 8*n+2 bytes.
1314 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1315 * At minimum we need 10 bytes (to generate 1 SACK block). If both
1316 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1317 * we only have 10 bytes for SACK options (40 - (12 + 18)).
1320 tcp_addoptions(struct tcpopt *to, u_char *optp)
1322 u_int mask, optlen = 0;
1324 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1325 if ((to->to_flags & mask) != mask)
1327 if (optlen == TCP_MAXOLEN)
1329 switch (to->to_flags & mask) {
1331 while (optlen % 4) {
1332 optlen += TCPOLEN_NOP;
1333 *optp++ = TCPOPT_NOP;
1335 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1337 optlen += TCPOLEN_MAXSEG;
1338 *optp++ = TCPOPT_MAXSEG;
1339 *optp++ = TCPOLEN_MAXSEG;
1340 to->to_mss = htons(to->to_mss);
1341 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1342 optp += sizeof(to->to_mss);
1345 while (!optlen || optlen % 2 != 1) {
1346 optlen += TCPOLEN_NOP;
1347 *optp++ = TCPOPT_NOP;
1349 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1351 optlen += TCPOLEN_WINDOW;
1352 *optp++ = TCPOPT_WINDOW;
1353 *optp++ = TCPOLEN_WINDOW;
1354 *optp++ = to->to_wscale;
1357 while (optlen % 2) {
1358 optlen += TCPOLEN_NOP;
1359 *optp++ = TCPOPT_NOP;
1361 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1363 optlen += TCPOLEN_SACK_PERMITTED;
1364 *optp++ = TCPOPT_SACK_PERMITTED;
1365 *optp++ = TCPOLEN_SACK_PERMITTED;
1368 while (!optlen || optlen % 4 != 2) {
1369 optlen += TCPOLEN_NOP;
1370 *optp++ = TCPOPT_NOP;
1372 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1374 optlen += TCPOLEN_TIMESTAMP;
1375 *optp++ = TCPOPT_TIMESTAMP;
1376 *optp++ = TCPOLEN_TIMESTAMP;
1377 to->to_tsval = htonl(to->to_tsval);
1378 to->to_tsecr = htonl(to->to_tsecr);
1379 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1380 optp += sizeof(to->to_tsval);
1381 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1382 optp += sizeof(to->to_tsecr);
1386 int siglen = TCPOLEN_SIGNATURE - 2;
1388 while (!optlen || optlen % 4 != 2) {
1389 optlen += TCPOLEN_NOP;
1390 *optp++ = TCPOPT_NOP;
1392 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE)
1394 optlen += TCPOLEN_SIGNATURE;
1395 *optp++ = TCPOPT_SIGNATURE;
1396 *optp++ = TCPOLEN_SIGNATURE;
1397 to->to_signature = optp;
1405 struct sackblk *sack = (struct sackblk *)to->to_sacks;
1408 while (!optlen || optlen % 4 != 2) {
1409 optlen += TCPOLEN_NOP;
1410 *optp++ = TCPOPT_NOP;
1412 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1414 optlen += TCPOLEN_SACKHDR;
1415 *optp++ = TCPOPT_SACK;
1416 sackblks = min(to->to_nsacks,
1417 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1418 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1419 while (sackblks--) {
1420 sack_seq = htonl(sack->start);
1421 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1422 optp += sizeof(sack_seq);
1423 sack_seq = htonl(sack->end);
1424 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1425 optp += sizeof(sack_seq);
1426 optlen += TCPOLEN_SACK;
1429 V_tcpstat.tcps_sack_send_blocks++;
1433 panic("%s: unknown TCP option type", __func__);
1438 /* Terminate and pad TCP options to a 4 byte boundary. */
1440 optlen += TCPOLEN_EOL;
1441 *optp++ = TCPOPT_EOL;
1444 * According to RFC 793 (STD0007):
1445 * "The content of the header beyond the End-of-Option option
1446 * must be header padding (i.e., zero)."
1447 * and later: "The padding is composed of zeros."
1449 while (optlen % 4) {
1450 optlen += TCPOLEN_PAD;
1451 *optp++ = TCPOPT_PAD;
1454 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));